Network State Estimation Via Passive Traffic Monitoring

We propose to study computer network traffic as a dynamical system, with the intent of determining how predictable the traffic is over short time scales. We will use passive measurements from high capacity links, so that we may investigate traffic that consists of many diverse component flows. To study network traffic as a dynamical system one must first have a concept of what variables compose the state space. Transmission Control Protocol (TCP) regulates the dynamics of flows through two primary variables, the round-trip time and congestion window. These are obvious choices for the state variables, but they are not recorded in the passive measurements. Our main contributions in this dissertation are two algorithms that estimate round-trip times and congestion windows, and an auxiliary algorithm that determines flow orientation. We provide several validation tests for the algorithms, and use the results of the algorithms to infer the level of congestion in the measurements we use